To Pay or be Paid? The Impact of Taker Fees and Order Flow Inducements on Trading Costs in U.S. Options Markets*

Transcription

1 To Pay or be Paid? The Impact of Taker Fees and Order Flow Inducements on Trading Costs in U.S. Options Markets* Robert Battalio Mendoza College of Business University of Notre Dame (574) Andriy Shkilko School of Business and Economics Wilfrid Laurier University (519) x2462 Robert Van Ness School of Business University of Mississippi (662) May 2012 Abstract: Equity options exchanges in the United States use one of two models to execute marketable orders: the maker-taker (MT) model or the payment for order flow (PFOF) model. Exchanges utilizing the MT model charge liquidity demanders a taker fee to access their liquidity. Exchanges utilizing the PFOF model use order flow payments to attract marketable retail orders. We examine whether the agency problems associated with PFOF manifest themselves in the equity options markets. Focusing solely on execution prices and controlling for the trading cost determinants, we find that the cost of liquidity on exchanges utilizing the PFOF model is 100 bps higher than on exchanges utilizing MT pricing. Nevertheless, when taker fees are incorporated into the analysis, the cost of liquidity on the PFOF exchanges is 40 bps lower. Our results have two implications: (i) transparency and competition in equity options markets appear to have limited the potential agency problems, and (ii) evaluations of market quality that ignore taker fees can be misleading. *We would like to thank Peter Bottini, Shane Corwin, Steve Crutchfield, Paul Gao, Maureen O Hara, Fabricio Perez, Tanseli Savaser, Paul Schultz, Jim Upson, Mao Ye and seminar/session participants at De Paul University, Florida International University, the University of Miami, the University of Notre Dame, Wilfrid Laurier University, Mid-Atlantic Research Conference in Finance, and Eastern Finance Association meetings for their comments. 0

2 1. Introduction Recent structural changes in the U.S. options markets have resulted in a unique trading landscape, whereby more than 70% of volume executes on exchanges that employ the Payment for Order Flow (PFOF) model. In this model, liquidity is supplied by market makers, who naturally prefer trading with the uninformed. To attract such orders, the PFOF exchanges administer a program whereby brokers get paid to route retail liquidity-demanding orders to the PFOF market makers. In the alternative maker-taker (MT) model, designated market makers play a trivial role in providing liquidity. Instead, liquidity is attracted by charging access fees to liquidity demanders and rebating a portion of these fees to liquidity suppliers the limit order traders. Markets using the MT model do not purchase retail liquidity-demanding orders. Brokers receive payments for retail orders from the PFOF markets, while having to pay to send the same orders to the MT markets. 1 Such a clear incentive to route to the PFOF venues is potentially problematic and may conflict with a broker s fiduciary duty to obtain the best execution for her clients. 2 Although national market rules protect the best quotes, PFOF arrangements may induce brokers to route to venues that offer less price improvement, lower speeds, and lower execution probabilities potentially resulting in inferior prices. Anecdotal evidence suggests that brokers in the options markets eagerly respond to the PFOF incentives. TD Ameritrade, for instance, annually generates $78 million from options PFOF arrangements. 3 Similarly, Interactive Brokers (IB) states that when multiple options exchanges are quoting at the 1 During our sample period, a broker pays up to $0.45 per contact ($ per option) to access MT quotes, whereas the PFOF markets pay the broker up to $0.70 per contact for retail orders. 2 In an October 2011 Proposed Guidance on the Practice of Payment for Order Flow, the U.K. s Financial Services Authority writes: [W]e believe PFOF arrangements create a clear conflict of interest between the clients of the [brokerage] firm and the firm itself. 3 See September 2010 GETCO letter to the SEC. 1

3 National Best Bid and Offer (NBBO), IB will generally send the order to an exchange where it will receive the most payment for the order. 4 Given the potential for agency conflicts induced by the fee structure in the U.S. options markets, it is important to assess execution cost differences between the PFOF and the MT models. Our data are uniquely suited for such assessment; we obtain a previously unexplored set of maker-taker fees and marketing fees (the latter are the fees paid by the PFOF market makers to fund order flow payments) from each of the major options exchanges. By incorporating these fees into the conventional measure of execution quality (i.e., effective spread), we not only expect to shed new light on the competitive landscape in the options marketplace, but also to enhance the extant literature that largely ignores the effect of access fees and order flow inducements on execution quality. We begin by comparing trading costs based on unadjusted effective spreads. When we control for conventional trading cost determinants, we find that an average liquidity demanding trader pays 100 bps more per option in the PFOF model than she pays in the MT model. This comparison, however, ignores the taker fees and therefore conceals a potentially significant portion of true costs. As noted by Angel, Harris, and Spatt (2011), if the market for order flow is sufficiently competitive, access fees are passed onto liquidity demanders either explicitly or, more likely, implicitly via higher commissions and/or reduced services. Once we incorporate taker fees into trading costs measures, the advantage of the MT venues suggested by the unadjusted spreads results fully reverses an average liquidity demanding trader pays 40 bps less per option in the PFOF model than she pays in the MT model. This result is important on two levels. First, it shows that the conventional cross-market comparisons of execution quality that often rely on unadjusted spreads should be interpreted with 4 2

4 caution. Second, it implies that the avoidance of taker fees rather than the receipt of order flow payments may explain why brokers seeking to obtain best execution for their clients rout to the PFOF exchanges. Next, we examine the potential effect of order flow inducements. Theoretical predictions of Battalio and Holden (2001) suggest that competition for retail order flow will force brokers to disburse the marketing payments obtained from the PFOF markets. We note that the payments may not be channelled to retail traders exclusively, especially given the current commission structure, whereby most options brokers charge a fixed commission to all of their customers. Instead, order flow inducements may be distributed among a wider set of brokerage clients. Based on this logic, we introduce a third measure of execution quality that accounts for the possibility that marketing fees evenly reduce trading costs for all trades in the PFOF markets. Because of the abovementioned commission rigidities, this measure overstates the effect of order flow inducements on trading costs, yet it highlights the cost advantage of the PFOF markets even further. Specifically, an average liquidity demanding retail trader pays up to 180 bps less to trade in the PFOF model than she pays in the MT model. We caution that our results do not suggest a clear advantage of the PFOF model over the MT model. It is quite likely that lower trading costs on the PFOF exchanges are attributable to the higher proportion of retail order flow routed to these exchanges. As such, our findings are consistent with what one might expect to find in a competitive and transparent market, in which exchanges specialize in the type of orders that they execute. The remainder of this paper is organized as follows. In the next section, we provide the institutional background and elaborate on the related literature. In Section 3, we describe our data and generate descriptive statistics. In Section 4, we present the results of our analysis of trading 3

5 cost differences between the PFOF and MT models. In Section 5, we discuss our findings. Section 6 concludes. 2. Institutional background and related literature The introduction of pennies to the U.S. equity option markets in 2007 coincides with the first use of the MT model by an options exchange. Prior to pennies, all exchanges used the traditional dealer-driven customer priority model (also known as the PFOF model), in which exchanges charge brokers little or nothing to execute customer orders, but charge liquidity providers marketing fees for trades with retail customer orders. These marketing fees are then used to pay for retail order flow. Unlike equity market makers, options market makers often do not execute all of the liquidity demanding orders they purchase. On PFOF exchanges, the percentage of the order that a purchasing market maker (PMM) can interact with depends on the number of competing market makers who are willing to match the PMM s price. For example, the CBOE s Automated Improvement Mechanism (AIM) allows its market makers to interact with at least 40% of purchased orders. Market makers executing orders via AIM must improve the NBBO by at least $0.01 (match the NBBO) if the marketable order is for less than 50 contracts (for at least 50 contracts). Programs on the Philadelphia Stock Exchange (the Directed Order Flow Program) and on the International Securities Exchange (Rule 713) allow market makers to interact with purchased orders in a similar fashion. Alternatively, in the MT model, the spread revenue is supplemented by rebates to orders that provide liquidity. These rebates are funded by charging a fee to orders that demand liquidity. As in equity markets, the fee structure on the options exchanges employing the MT model 4

6 provides incentives for market participants to quote aggressively, which attracts order flow. While MT exchanges typically follow price-time priority, exchanges using the PFOF model generally give customer orders priority at the best price over other trading interests at that price. The MT model seems to have found a home in equity options markets, and its popularity notably increased in recent years. The percent of option contracts traded on exchanges using MT pricing increased from 12% in the first quarter of to more than 22% in our 2010 sample. What are the implications of the two competing trading models for liquidity demanding investors? In this paper, we examine whether PFOF exchanges offer inferior execution prices relative to the execution prices on MT exchanges. Given the magnitude of order flow inducements and the dominant market share of the PFOF exchanges, the options market gives us an ideal opportunity to identify the adverse effects of order flow inducements. Despite convincing evidence that order flow routed pursuant to PFOF arrangements is less informed, research has not arrived at a consensus as to whether uninformed retail investors are harmed by these arrangements. Early equity market studies examining trade and quote data (e.g., Lee, 1993; Petersen and Fialkowski, 1994; Easley, Kiefer, and O Hara, 1996; Bessembinder and Kaufman, 1997) suggest the orders routed pursuant to PFOF arrangements pay higher effective spreads than comparable orders. Meanwhile, studies by Battalio (1997) and Battalio, Greene and Jennings (1997, 1998) show that the introduction of market makers who respectively purchase, preference, and internalize orders in the NYSE-listed stocks is associated with a tightening of the National Best Bid and Offer (NBBO). Battalio and Holden (2001) argue that net costs should be considered when evaluating execution quality and show that, depending on the amount of payment that ultimately reaches the 5 See "BATS Aims to Mirror Equities Success in Options," by Nina Mehta in the July 10, 2009 edition of the Traders Magazine. 5

7 retail investor, brokers who route orders pursuant to payment for order flow agreements may offer lower net trading costs. Battalio, Jennings, and Selway (2001) show that the net cost of liquidity offered by some brokers who sell orders in NASDAQ stocks to Knight Securities is lower than the net cost of trading through Trade Fast, the only broker in their sample that does not receive inducements for its equity order flow. In equity markets, Barclay et al. (2003) and Goldstein et al. (2008) find that electronic trading venues that use the MT model are associated with tighter quoted spreads and, in general, lower effective spreads. We extend this literature by examining the impact of MT exchanges in options markets and by conducting a comprehensive analysis of the impact of taker fees on trading costs. Our results confirm the abovementioned studies in that MT venues often offer lower unadjusted spreads, yet this result reverses once we account for taker fees. On the feeadjusted basis, traders generally pay more to execute through an MT market. More recently, O Hara and Ye (2011) find that increased fragmented trading in the U.S. equity markets circa 2008 is associated with lower transaction costs, faster execution speeds, and greater market efficiency. On the contrary, Weaver (2011) focuses on fragmentation caused by internalization and finds it to be associated with wider quoted and effective spreads. Our study complements this research by examining how two factors that influence order routing decisions order flow inducements and taker fees affect execution quality. Finally, Malinova and Park (2011) examine the Toronto Stock Exchange s decision to change from a value to a volume-based fee structure. Their findings support the predictions by Colliard and Foucault (2011) that the bid-ask spread decreases when the take fee or the liquidity rebate increases. We extend this strand of literature by employing cross sectional analysis to 6

8 examine measures of gross and net trading costs in a market where exchanges utilizing the MT model are in competition with those employing the PFOF model. As a by-product of our research, we find that the execution quality statistics currently published monthly on a voluntary basis by the seven options exchanges are often unreliable. When we compare the exchange-generated effective spreads with spreads that we compute using publicly available trade and quote data, we find large discrepancies for five out of seven exchanges. These five exchanges produce their numbers internally. Conversely, our monthly average spread estimates are consistently close to those produced by an external firm, S3, 6 for the AMEX and for the NYSE Arca. Based on this finding, we conduct our analysis using relative effective spreads generated using public trade and quote data rather than using the exchangegenerated statistics. 3. Data and descriptive statistics 3.1 Sample structure and filters Options exchanges report trade and quote data via the Options Price Reporting Authority (OPRA). In turn, OPRA provides price and volume information on trades and on current bids and offers in eligible securities from 7:30 a.m. to 6:00 p.m. Eastern time on any regular trading day (see SEC Release No ). Using OPRA data, a data technology company LiveVol creates historical data files that contain information on every option trade each day, including the bid and ask quotations prevailing at each of the eight options exchanges when the trade is reported to OPRA. 7 We obtain historical data from LiveVol for a period from March 1, 2010 through June 30, Due to data corruption issues, we drop March 22 and March 25 trading

9 days. In addition to options data, we use TAQ and CRSP data to compute regressors for the multivariate tests. Our initial sample contains trade and matched quote data for 3,233 option classes. These classes trade on eight options exchanges: NASDAQ, AMEX, CBOE, ISE, NYSE Arca, PHLX, BOX, and BATS. Among these, the AMEX, the CBOE, and the ISE use the PFOF model; NASDAQ, the BOX, and BATS use the maker-taker model; and the NYSE Arca and the PHLX use a mix of the two models. To ensure proper comparison among exchanges, we restrict the sample to option classes that trade on all exchanges other than BATS. 8 This screen retains 550 option classes. After eliminating option classes that have fewer than 10 trades per day, we are left with 314 classes. We next exclude options on foreign stocks, ADRs, and REITs, and option classes that switch to trading in pennies during our sample period. Our final sample consists of 239 option classes, with 32 option classes on ETFs and 207 classes on common stocks. Out of 239 classes, 122 trade in pennies during the entire sample period (with 27 out of the 32 ETF option classes trading in pennies). Panels A and B of Table 1 contain summary statistics for our data filters and the resulting sample. [Insert Table 1 about here] LiveVol trade records contain the date and to-the-second time, option class, strike price, expiration date, put/call indicator, trade price, trade size, trade condition identifier, National Best Bid (NBB) price and size, and the National Best Offer (NBO) price and size prevailing when the trade is executed, bid price, bid size, offer price, and offer size prevailing at each of the eight exchanges when the trade executes, and the underlying stock s NBBO when the trade executes. 8 Our decision to exclude BATS is due to its trivial (less than 1%) share of contract volume during the sample period. 8

10 At any point in time, the NBB is the highest bid price and the NBO is the lowest offer price across the eight option exchanges. As noted by Battalio, Hatch and Jennings (2004), it is common for orders arriving on PFOF exchanges to be divided among all counterparties willing to match the best available price. Similarly, orders arriving at maker-taker exchanges may execute against multiple liquidity suppliers at a given price. Thus, a single order often produces several trade records. We follow Battalio et al. (2004) and combine multiple executions in the same option series, executing on the same exchange at the same price at the same time with the same trade condition identifier into a single trade. As noted in Panel C of Table 1, our initial sample contains over 21 million trades. To avoid trades involved in the opening and closing rotations, we eliminate trades reported before 9:45 a.m. and after 3:55 p.m. each day. Across the eight exchanges in our sample, this screen eliminates 11.27% of reported trades. Since complex trades (e.g., spreads and straddles) are priced as a package, we eliminate them too. Three of the maker-taker exchanges trade no spreads or straddles during our sample period, while the ISE reports that over 35.25% of its trades are spreads or straddles. Overall, this screen eliminates about 10.48% of the sample trades. We next eliminate trades with benchmark execution-time NBBOs that are crossed and trades with benchmark execution-time NBBs that are equal to zero, reasoning that the execution-time NBBO may not be an appropriate execution quality benchmark for these trades. Together, the crossed markets screen and the zero NBB screen eliminate less than 0.5% of the sample. After applying these screens and then bunching trades as described in the previous paragraph, we are left with a little more than 61% of the trades in our initial sample for a total of nearly 13 million trades. 9

11 Next, we turn our attention to the distribution of trading activity among sample venues. Panel A of Table 2 contains statistics on the percentage of trades and volume while grouping the sample venues into pure PFOF, mixed, and pure maker-taker models. The three exchanges that operate solely under the PFOF model (i.e., CBOE, ISE, and AMEX) execute 56.22% of the trades and 58.54% of the contract volume. Their market share is around 55% of contract volume for ETF options and over 63% of contract volume for options that do not trade in pennies. The CBOE has the dominant market share, whereas the AMEX and the ISE execute roughly the same percentage of trades. [Insert Table 2 about here] Panel B of Table 2 reports the percentage of class days during which each sample exchange is the dominant market and the percentage of class-days during which each exchange has no trading activity in our sample. Consistent with their overall market share, the CBOE and the ISE are the dominant markets for the majority of trading days. One of these two exchanges is the dominant market during over 61% of class-days. On 13.35% of class-days, the NYSE Arca is the dominant market, while the pure maker-taker markets have the dominant market share during only 1.7% of the class-days. BATS is never the dominant market and has no trading activity during about 89% of class-days. NASDAQ and the BOX have no trading activity for 7.2% and 3.7% of the class-days, respectively, while the NYSE Arca is inactive for 2.5% of the class-days. The traditional PFOF exchanges each have no trading activity for less than 0.3% of the class-days. Panel C of Table 2 describes the distribution of trade sizes across exchanges. The exchanges that solely employ the PFOF model execute over 55% of the trades for up to 50 contracts. These trades most likely represent the execution of retail trading interest. The 10

12 aggregate market share for these exchanges jumps to about 63% for trades of more than 50 contracts. This increased market share may be attributable to depth guarantees made by the PFOF exchanges to the retail brokers from whom they purchase orders and to the fact that liquidity providers generally do not have to improve upon the NBBO to interact with large orders. The pure maker-taker exchanges execute more trades between 2 and 50 contracts (12.26%) than they do trades of 1 contract (8.26%) or trades of more than 50 contracts (5.25%). The NYSE Arca s market share is greatest in 1 contract trades (20.97%) and falls to a market share of 14.70% for trades of more than 50 contracts. Finally, the PHLX executes 14.02% of trades for 2 to 50 contracts and just over 17% of trades for more than 50 contracts. In Figure 1, we provide an illustration of the daily share of contract volume executed on the PFOF exchanges during our sample period. The percentage of contracts that execute on exchanges using the PFOF model declines from just under 77.5% on March 1, 2010 to just under 71.4% on June 30, 2010, perhaps reflecting the fact that the PHLX moves many of its actively traded option classes from the PFOF model to the maker-taker model during our sample period. There is a similar decline in the PFOF market share in the penny option classes over our sample period; the PFOF exchanges market share peaks at 67.8% on March 10 and reaches a low of 55.7% on May Overall, in stark contrast to equity markets, option exchanges offering order flow inducements to liquidity demanders have a dominant market share. [Insert Figure 1 about here] 3.2 Fees and rebates The goal of this study is to compare the net execution costs between the PFOF and MT models. We therefore obtain information on taker fees and maker rebates (Panel A of Table 3) 9 Perhaps surprisingly, we do not observe significant migration of trades from the maker-taker exchanges (the exchanges without traditional market makers) to the PFOF exchanges (the exchanges with traditional market makers) on May 6, the day of the flash crash. 11

13 and on marketing fees (Panel B) for each exchange other than BATS from a major options market maker. As suggested in the SEC s April 2010 rule proposal, there is a clear bifurcation in the data; each exchange trades a specific option class under either the maker-taker model or the PFOF model. 10 [Insert Table 3 about here] As indicated in Panel A, the CBOE, the ISE, and the AMEX do not charge maker-taker fees. On each of these exchanges, market makers pay a fee of $0.25 per contract ($0.65 per contract) when they provide liquidity to retail customers trading options (not) participating in the penny pilot (Panel B). Conversely, NASDAQ and the BOX do not charge marketing fees during our sample period, but they do charge maker-taker fees. In penny pilot options, taker fees on NASDAQ are $0.35 per contract and maker rebates are $0.25 per contract. The BOX has a nominally inverted fee structure, whereby it seemingly charges liquidity providers in penny options $0.15 per contract and rebates this amount to liquidity demanders. For the purposes of our analysis, the fee structure on the BOX is more conventional that it may seem at first glance. The BOX is rarely at the NBBO, and therefore most liquidity-taking orders routed to it are subject to the Price Improvement Period (PIP). 11 During PIP, these orders are flashed for an instant and effectively become short-lived liquidity-making orders. If traded against, these orders earn the taker rebate. The orders that respond to the flashed orders are charged a maker fee, which most readers would qualify as a conventional charge for interacting with a sitting order. Thus, despite the unconventional terminology, liquidity rebates on the BOX effectively go to liquidity providers, and liquidity demanders are charged taker fees. 12 All things 10 See SEC Release , Proposed Amendments to Rule 610 of Regulation NMS, dated April 14, Several industry insiders suggested to us that the BOX chooses to specialize in internalized order flow. When an internalizing trader submits a liquidity demanding order to the BOX, the exchange initiates a PIP, during which 12

14 considered, the fee structure on the BOX is rather conventional. All results presented in the subsequent sections are robust to excluding trades executed on the BOX. The remaining two exchanges do not exclusively use the MT or the PFOF model. The NYSE Arca uses the PFOF model when trading options not participating in the penny pilot, and it uses the maker-taker model for penny pilot options. The PHLX began using the maker-taker model for 27 actively traded option classes on March 1, 2010, then added an additional 24 actively traded option classes on April 13, 2010, 4 actively traded option classes on May 6, 2010, and 23 actively traded option classes on June 4, The NYSE Arca charges taker fees of $0.45 per contract and offers rebates of between $0.25 and $0.30 per contract. The PHLX charges taker fees of $0.25 per contract and offers rebates of $0.20 per contract. For those option classes trading under the PFOF model, the NYSE Arca charges marketing fees of $0.65 per contract, while the PHLX charges $0.25 per contract ($0.70 per contract) for options (not) trading in the penny pilot program. In the remaining analysis, we classify trades as occurring either on an exchange utilizing the PFOF model or the MT model. We begin by examining whether there are systematic differences in execution costs between the two models. We first use raw relative effective spreads as our summary measure of execution quality. Next, we adjust relative effective spreads to reflect the taker fees paid on exchanges utilizing the MT model. Our final measure of execution quality assumes that the entire marketing fee paid by liquidity suppliers on PFOF exchanges is paid to brokers and is then passed onto liquidity demanders. market participants may respond to the PIP order by submitting Improvement Orders. The fee for liquidity provision however deters them from responding, and often the originating internalizing trader is the only one to respond. For this trader, the liquidity provision fee cancels out the liquidity demand rebate as the trader successfully internalizes his entire order. 13

15 As an alternative to the abovementioned measures, we considered using order-based execution quality statistics for market and marketable orders that are voluntarily published by all options exchanges other than BATS. These statistics may be more suitable for our analysis, as they (i) are benchmarked against order receipt time quotes and (ii) mainly focus on retail customer orders. Nevertheless, a careful analysis of these statistics indicates that they are not of sufficient quality for five out of seven trading venues. In the meantime, we are encouraged that the statistics published by the two exchanges (the NYSE Arca and the AMEX) that outsource computations to a well-respected third party are noticeably comparable to our trade-based statistics. We discuss this issue in detail in Appendix I. 4. Execution quality on maker-taker and PFOF exchanges 4.1 Univariate analysis of execution quality Unadjusted relative effective spreads. To examine if the PFOF markets offer inferior executions, we compute, for each trade in our sample, the relative effective spread as twice the absolute difference between the trade price and the execution time NBBO midpoint divided by the midpoint. We then compute equal-weighted spreads as follows. Each trading day, for each option class, we compute an average of the relative effective spreads of the underlying option series. We then compute the average spread across all option classes on that day. In columns 1 and 3 of Table 4, we report the equal-weighted time series averages of raw relative effective spreads for PFOF and MT markets. On average, raw spreads are 80 bps lower on the MT exchanges than on the PFOF exchanges. This is despite the fact that the PFOF exchanges generally receive order flow that is less informed. 14

16 As we show in Panel A of Table 2, trades on the NYSE Arca make up the majority of trades in the MT model. To ensure that our results are generalizable beyond the NYSE Arca, we next present relative effective spreads for the MT and the PFOF exchanges excluding trades on the NYSE Arca. Consistent with the overall results, relative effective spreads on the MT markets are 47 basis points lower than those on the PFOF exchanges. We obtain qualitatively similar results when we exclude the PHLX trades and when we exclude the BOX trades. In Appendix II, we show that these results hold when we slice the data according to trade size, equity vs. ETF options, and penny vs. non-penny options. Further, the results remain qualitatively unchanged when we compare the PFOF and MT models within the NYSE Arca and the PHLX. Finally, the multivariate framework discussed in the next section allows us to control for a number of characteristics (including the abovementioned) that may potentially affect trading costs. [Insert Table 4 about here] It is possible that our results are affected by quote competition between the two models. Specifically, MT markets may provide liquidity when NBBOs are narrow and leave liquidity provision to PFOF markets when NBBOs are wide. To examine this possibility, we next restrict the sample to trades executed when at least one PFOF and at least one MT market are at the NBBO. Our results remain qualitatively unchanged although the difference between the PFOF and MT models declines; unadjusted spreads in the PFOF markets are now only 41 bps higher. In the multivariate analysis that follows, we further examine how quote competitiveness affects trading costs in our setting. Relative effective spreads adjusted for taker fees. Angel et al. (2011) note that to earn liquidity rebates, liquidity providers on MT exchanges tend to offer better prices, which leads to tighter bid ask spreads. In competitive markets, the taker fees charged by the MT venues should 15

17 offset the tighter spreads on these venues so that the cost of liquidity net of the taker fees is the same on PFOF and MT markets (Colliard and Foucault, 2011; Malinova and Park, 2011). This argument suggests that there should be no difference in the effective spreads on PFOF and MT exchanges after fees are taken into account. In what follows, we test whether this conjecture holds in our sample. We begin by computing the relative round trip taker fee for each trade on an MT exchange. We then divide the taker fee by 100 so that it is on a per contract basis. We next multiply the adjusted taker fee by two so that it reflects the per contract fee paid on a round trip liquidity demanding trade. We then divide this amount by the midpoint of the execution-time NBBO. We arrive at our fee-adjusted relative effective spread (column 4 of Table 4) by adding this amount to the trade s relative effective spread. When we ignore taker fees, the relative effective spread is about 80 bps lower on the MT exchanges. When we incorporate the fees, this differential declines, and PFOF spreads are higher by an average of only 14 basis points. When trades on the NYSE Arca are excluded from our analysis, results are quite similar to those for the entire sample, yet the differential is larger, at 40 bps. The differential becomes 20 bps when we exclude trades on the BOX. When we exclude PHLX and when we focus solely on trades executed when both models are at the NBBO, the differential between the two models disappears. Relative effective spreads adjusted for taker fees and order flow payments. Battalio and Holden (2001) argue that if the brokerage industry is competitive, and if payments are transparent, competition will force brokers to pass order flow inducements to their customers in the form of lower brokerage commissions. Because the SEC Rule 606 requires brokers to 16

18 disclose information regarding the order flow inducements they receive, the net cost of trading should be lower on PFOF exchanges if there is sufficient competition in the brokerage industry. We next examine relative effective spreads adjusted for both taker fees and marketing fees under the assumption that 100% of each of these fees is passed on by brokers to retail liquidity demanders. 13, 14 Unlike taker fees, marketing fees will reduce the overall cost of liquidity. We first compute a relative round trip order flow payment for each trade on a PFOF exchange by dividing the marketing fee by 100 so that it is on a per contract basis. We next multiply the adjusted marketing fee by two so that it reflects the per contract order flow payment made on a round trip liquidity demanding trade. We then divide this amount by the midpoint of the execution-time NBBO. We arrive at our payment-adjusted relative effective spread by subtracting this amount from the trade s relative effective spread. Equal-weighted paymentadjusted relative effective spreads are presented for the PFOF exchanges in column 2 of Table 4. Our findings illustrate the importance of order flow inducements and taker fees (if passed on to investors) on inferences as to which type of market offers the cheapest liquidity. Taken at face value, the results in Table 4 do not imply that MT exchanges offer non-competitive execution prices. Rather, these results likely reflect the fact that MT exchanges receive only those orders that cannot be sold to the PFOF exchanges. Presumably, these orders are more informed and therefore costlier to execute. In a competitive market, one would expect such orders to pay higher unadjusted effective spreads. 13 Angel et al. (2011, p. 23) note that since brokers cannot obtain payments (order flow inducements) if they do not have retail orders, competition forces the brokers to return much, if not all of these payments to their clients in the form of lower commissions or better services, both of which attract retail clients and their orders. 14 This assumption presumes that there is no cross-subsidization across a broker s clientele (e.g., market order traders versus limit order traders, professional traders versus retail investors). 17

19 4.2 Multivariate analysis of execution quality Options exchanges typically trade hundreds of different options on a given stock. Some of these options are liquid, while others are not. Some trade in pennies, while others trade in nickels and dimes. Thus, even though we require our sample option classes to average at least ten trades per day, there is no guarantee that the liquidity of options that execute on each exchange is comparable. For example, if certain exchanges specialize in providing executions for the most liquid options, simple univariate comparisons may lead us to incorrectly conclude that these exchanges offer superior execution quality. Similarly, if markets in one model usually receive order flow that is more toxic, our univariate results may incorrectly identify them as providing inferior executions. 15 For this reason, we next examine whether our univariate results survive when we control for possible selection biases in the following panel regression model: Since trades represent only one side of a round-trip transaction, for each trade in option class i executed at time t we measure the dependent variable,, as twice the absolute difference between the trade price and the midpoint of the execution-time NBBO normalized by the midpoint. The variable of interest is, which equals one if the trade occurs in the PFOF model and zero if the trade occurs in the MT model. Since the $0.05 minimum variation creates a lower bound on effective spreads for options that do not trade in pennies, we include an indicator variable that is equal to one if the 15 See Bessembinder (2003b) for a discussion of selection biases in multi-market studies. 18

20 option trades in pennies and is equal to zero otherwise. To control for potential differences in hedging costs in options on ETFs and options on equities, we include an indicator variable that is equal to one if the option is on an ETF and is zero otherwise. Given the well documented relationship between spreads and volatility, equation (1) includes, the trade s implied volatility as computed by LiveVol. Further, theoretical predictions by Easley and O Hara (1987) suggest a positive relation between liquidity costs and trade size, therefore we include the natural log of the trade size,, as a control. In addition, since liquidity is a function of an option s moneyness, we include a dummy variable that equals to 1 if the option s stock-to-strike price is between 0.9 and 1.1. We also include a dummy variable that equals to 1 if the option s stock-to-strike price is either between 0.7 and 1.1 or is between 1.1 and 1.3. Additionally, we include the execution-time NBBO midpoint,, since prior work (e.g., Harris, 1994) shows that spreads depend on price. Following Chakravarty, Jain, Upson, and Wood (2011), who show that informed traders often use Intermarket Sweep Orders (ISO), we introduce the indicator variable. 16 To account for possible differences between call and put options, we use a binary indicator. Unlike equity market makers, option market makers quickly hedge newly opened positions in the underlying stock (Jameson and Wilhelm, 1992; Battalio and Schultz, 2011). Therefore, researchers do not usually use realized spreads to control for selection bias related to trades information loads. In addition, a number of option series trade rather infrequently and computing high-frequency price impact measures for them is impractical. It may be argued that the movement in the underlying stock price following a trade in the option market might be used as a proxy for the information content of a liquidity demanding option trade. There are two 16 In our sample, close to 6% of trades originate from ISO orders. 19

21 problems, however, with this proxy. First, the proxy will not control for volatility option trades which, by definition, are non-directional. Second, for directional trades, it is difficult if not impossible to distinguish between movements in the underlying stock that reflect information versus movements that reflect the hedging trades of options market makers. We recognize that it is important to keep hedging costs constant. Order flow toxicity may affect trading costs in the underlying equity and thereby option market makers hedging costs. To proxy for hedging costs, we use effective spreads in the underlying equity market lagged by one second,. In addition, to control for the possibility that either PFOF or MT markets receive more order flow when option spreads are wide, we keep the NBBO constant. Furthermore, following Battalio et al. (2004), we include the natural log of the underlying asset s market capitalization,, computed each day using CRSP data. is intended to capture the effect of long-run information asymmetries and liquidity in the underlying equities. To capture option expiration day effects, we include a binary variable that equals one for trades that execute on March 19, April 16, May 21, and June 18. We also include binary indicator variables and that equal to one for trades that execute, respectively, on May 6, 2010 (the day of the Flash Crash) and after May 6. Given that exchange traded funds are more liquid than stocks, we expect effective spreads to be smaller for options on ETFs. We also expect spreads to be smaller for options that trade in pennies. In addition, relative effective spreads should be increasing in option s implied volatility and in trade size and decreasing in moneyness and the market capitalization of the underlying stock. Underlying equity trading costs and the option s own NBBO should have a positive effect on spreads. We expect spreads to be higher for trades originating from ISO orders 20

22 and for trades executed on expiration days and on the day of the Flash Crash. To the extent that the effects of the crash were persistent, we expect spreads to be wider in the post-crash period. We estimate the model with an adjustment for option class fixed effects and include indicator variables for 30-minute intraday periods to control for intraday effects. In addition, we allow standard errors to cluster at the option class level. Option class fixed effects prevent us from computing the coefficients of the and indicators, because these indicators do not vary within classes. Given our interest in the effects of these two variables, we estimate eq. (1) twice: with and without class fixed effects. Finally, to gauge the effects of taker fees and order flow inducements, we estimate eq. (1) for the following three variants of the dependent variable: (i) unadjusted; (ii) adjusted for taker fees; and (iii) adjusted for both taker fees and marketing payments. Our results for the entire sample are presented in Panel A of Table 5. [Insert Table 5 about here] The first and second columns of Table 5 contain OLS and OLS with fixed effects regression results when eq. (1) is estimated using unadjusted spreads. Most of the results are virtually identical between the two estimation approaches, and we focus on the results without fixed effects. Consistent with our expectations, relative effective spreads increase in implied volatility and trade size and decline in option moneyness. Spreads are higher for trades that originate from ISO orders and are higher on expiration Fridays. The cost of liquidity as measured by relative effective spreads is 180 bps higher on the day of the Flash Crash and is 40 bps higher in the post-crash period. As expected, trading costs are higher when effective spreads in the underlying equities are large, and when the options own NBBOs are wide. Relative effective spreads are significantly lower for options that trade in pennies (around 330 basis points). Against our expectations, market capitalization and the ETF status of the underlying do not have 21

23 a significant effect on options trading costs. Also curiously, calls are somewhat more expensive to trade than puts. Full-sample univariate analyses in Table 4 show that the unadjusted trading costs for trades that execute in the PFOF model are 80 basis points higher than in the MT model. The multivariate analysis in Table 5 suggests that this difference is somewhat understated. The estimated coefficient of the PFOF dummy is (t-statistic = 13.61) in the no-fixed effects regression and is (t-statistic = 18.15) in the fixed effects regression. Thus, liquidity demanding trades that execute in the PFOF model appear to pay 100 bps more than in the MT model. The third and the fourth columns of Table 5 contain the results obtained when we use the fee-adjusted relative effective spread as the dependent variable. The inclusion of taker fees clearly negates the cost advantage of the MT model. When we assume that investors pay 100% of the taker fees, the results show that liquidity demanding trades in the PFOF model pay adjusted spreads that are 40 bps lower than those that execute in the MT model. Finally, the fifth and the sixth columns of Table 5 contain results obtained when we use the fee- and payment-adjusted trading costs as dependent variable. Consistent with our univariate analysis, the assumption that 100% of both taker fees and marketing payments flows through to the liquidity demander further reduces the relative cost of liquidity on the PFOF exchanges. Our estimates suggest that the relative trading costs are now 180 bps lower on the PFOF exchanges. To make sure that our results are not driven by the fee structure on the BOX, we reestimate eq. (1) excluding BOX trades. The coefficients of the variable reported in Panel B of Table 5 are notably similar to those obtained for the full sample. 22

24 Bessembinder (2003c) shows that quote competition often leads to improved execution quality. Although eq. (1) keeps options NBBO constant, we recognize quote competition may have a more nuanced effect on our results. In our next step, we augment eq. (1) by adding two quote competition variables as follows:, where, and are binary indicators that capture instances during which a PFOF market or an MT market are at the best quote (best ask for liquidity demanding buys, and best bid for sales). The intercept absorbs instances whereby both models are at the best quote. This analysis requires us to sign trades, and we do so by assuming that trades executing above (below) the NBBO midpoint are initiated by buyers (sellers). We exclude trades that execute at NBBO midpoints, which reduces the size of our sample by 8.44% as compared to the sample used in Panel A of Table The remaining variables are defined as in Table 5. We report estimation results for eq. (2) in Panel A of Table The results suggest that when a PFOF market is alone at the best quote, unadjusted spreads are 70 bps lower, fee-adjusted spreads are 60 bps lower, and fee- and payment-adjusted spreads are 20 bps lower than they are when both models are at the best quote. Notably, when an MT market is alone at the NBBO, only the unadjusted spreads are significantly lower, by 60 bps, and both variants of adjusted 17 Trade classification algorithms such as that of Lee and Ready (1991) are not always suitable in the options market given infrequent trading of some option series. Our results are however robust to using the Lee and Ready algorithm. 18 To conserve space, we do not report the control coefficients unrelated to quote competition and model type in Panels A and B. A comprehensive set of coefficients is available upon request. 23

25 spreads are unaffected. These results are consistent with quote competition through quote improvement. That being said, the statistical and economic significance of the indicators remains virtually unchanged. Unadjusted PFOF spreads are 90 bps higher, fee-adjusted PFOF spreads are 30 bps lower, and fee- and payment-adjusted PFOF spreads are 190 bps lower than their MT counterparts. [Insert Table 6 about here] In Panel B of Table 6, we further restrict the sample to trades that execute when both a PFOF and an MT market are at the NBBO. This restriction reduces the sample size from about 13 million trades to 3.5 million trades. We report the estimates of the coefficient in Panel B of Table 6. Whereas the difference between the PFOF and MT markets remains at 140 bps for unadjusted spreads, the difference disappears once we adjust for taker fees. When we adjust for marketing payments, trading costs in the PFOF model become 170 bps lower than they are in the MT model. There may be some concern that our results simply reflect differences across exchanges. We address this concern in two ways. First, we replace the indicator in eq. (1) with a binary indicator for each exchange. For this analysis, we invert the maker-taker fee structure on the BOX to be consistent with our earlier description of liquidity provision on this venue. Consistent with our prior results, we find that the (un)adjusted relative effective spreads are (higher) lower on each of the pure PFOF exchanges than they are on each of the pure MT exchanges. This suggests that our results are not being driven by a single exchange. Second, we estimate model (1) separately for trades on the two exchanges that utilize both the PFOF model and the maker-taker model the NYSE Arca and the PHLX. Unfortunately, the NYSE Arca uses the MT model for options that trade in pennies and the 24

26 PFOF model for options that do not trade in pennies. For this reason, we cannot distinguish the incremental impact of trading in pennies (not trading in pennies) from the incremental impact of trading in the MT model (the PFOF model) for the NYSE Arca trades. We do not have this problem when analyzing trades on the PHLX. The results of these regressions are presented in Table 7. [Insert Table 7 about here] When significant, each of the control variables has the predicted sign when we run each of the three variants of our fixed effects model using only the NYSE Arca trades and only the PHLX trades. While we recognize that we cannot identify the impact of the PFOF and MT models on the NYSE Arca s trading costs, we present the NYSE Arca results for completeness. As is the case for the entire sample, the coefficient on the indicator variable declines as spreads are adjusted first for taker fees and then for both taker fees and order flow inducements. Consistent with the overall results, trades that execute via the PFOF model on the PHLX have unadjusted relative effective spreads that are, on average, 150 basis points higher (t-statistic of 6.51) than comparable trades that execute via the MT model (see column 2 of Table 7). After accounting for taker fees, there is no significant difference in fee-adjusted spreads on the PHLX. Finally, when both taker fees and order flow inducements are considered, net trading costs are significantly lower for trades that execute via the PFOF model. To summarize, the results obtained using trades on a single exchange that utilizes both models are consistent with those obtained from trades originating on exchanges that use a single model. This suggests that our results are perhaps not the product of unidentified exchange characteristics. 25